Research And Implementation On Models Of Profile-guided Compilation

Compiler optimization is a very important task which attemps to improve objectCode so as to make full use of performance of target machine. Traditional compileroptimizing techniques are mostly based on static analysis of source code, and allinformation needed by optimizer are collected staticlly. Such static information areconservative and often makes optimizer lose quite lot of oppommities for optimization.The profiling-guided compilation guides optimizing by run time profilinginformation. This method can remedy the disadvantage of static analysis method andgreatly improve the effct of optimizing and the execution efficiency of object code,which is impossible to achieve by traditional static methods. Profiling-guided compilingmethod has widely adopted by morden advanced compilers.In this paper, we discuss the issues about profiling-guided compiling techniquesand its implementation method from following aspects:We gave a brief introduction to the principle of profiling-guided compiling, theproblems we might face and solutions to them.We analyzed categories of information needed and collected by profiling-guidedcompiling, including their definition, application and how to capture them.We deeply analyzed the implementation method used by GCC4.0 such as thecompiling frame, the method of instrument, the format of profiling data file, and how toget profiling information and noting them in the RTL.We implemented a new profiling, path profiling, in GCC4.0 by use of thearithmetic reported by T. Ball and J. R. Larus. Path profiling record the times each pathwas executed and has lots of potential uses in programs optimizing and in softwaretesting, the experiment show that the method has achieved success.